LiveKit Agents + MemorySync — Long-Term Memory Example

A runnable LiveKit Agents voice agent with long-term memory backed by MemorySync. Tell the agent something, hang up, call back — it remembers.

Voice is the one surface where memory latency is audible, so the integration is built around a hard rule: recalled context is injected in on_user_turn_completed under a strict time budget (default 1.2 s). If the memory service is slow, the agent replies without memories — the conversation is never stalled. A background prefetch warms the next turn's recall, so the steady-state cost is close to zero.

• Add a .env in this directory with your LiveKit and MemorySync credentials:

  LIVEKIT_URL=your_livekit_url
  LIVEKIT_API_KEY=your_api_key
  LIVEKIT_API_SECRET=your_api_secret
  MEMORYSYNC_API_KEY=your_memorysync_api_key
  

  Get a MemorySync API key at app.memorysync.io (Settings → API Keys).
• Install dependencies:

  pip install -r requirements.txt

python memorysync_memory.py console

Say "My name is Alex and my favorite color is teal", exit, then start a new console session — the agent recalls both facts.

• memory.on_user_turn(turn_ctx, new_message) runs inside on_user_turn_completed and injects a compact block of relevant memories for this turn only (it is not persisted into the LLM context, avoiding context bloat and double-learning).
• memory.attach(session) subscribes to conversation_item_added and stores both user and assistant turns as they finalize, with deterministic idempotency seeds so retries and reconnects never create duplicates. Interrupted assistant turns are stored with interrupted: true metadata.
• create_memory_search_tool(memory) gives the LLM an explicit search tool for questions like "what did I tell you last week?". Errors come back to the model as readable strings, never exceptions.
• Memory outages, quota limits, and dead networks all degrade to "no memories this turn" — the call itself is never affected.

With speech-to-speech realtime models, prefer the memory search tool over per-turn injection: on_user_turn_completed still fires, but injection can land after the model has started speaking.

• MemorySync LiveKit guide — full integration reference
• livekit-memorysync on PyPI